Abstract:
Optical and chemical images are collected from ice cores and are useful for the ice science community to study the evolution of paleoclimate. In the first part of this thesis, an exploratory analysis of cleaning visual images is carried out. Visual images need some pre-processing cleaning steps in order to remove artifacts and prepare the images to be segmented and features in the ice, such as grain boundaries, isolated. The second part of the work deals with image registration between chemical and visual images. Two deep learning models, LKU-net and VoxelMorph, are implemented to register pairs of visual-chemical images. The proposed models are then compared with other registration methods to evaluate their performances and are also tested against noisy segmentations to see how performance degrades as noise increases.
